Grammaticality, Acceptability, and Probability: A Probabilistic View of Linguistic Knowledge.

Area of Science:

• Cognitive Science
• Linguistics
• Psychology

Background:

• The nature of grammatical knowledge (binary vs. probabilistic) is a long-standing debate.
• Acceptability judgments pose challenges for both binary and probabilistic grammaticality theories.
• Sentence acceptability is distinct from occurrence probability due to factors like length and frequency.

Purpose of the Study:

• To investigate the pervasive gradience in acceptability judgments.
• To develop a method for predicting acceptability judgments using probabilistic language models.
• To explore the implications for the debate on grammatical competence.

Main Methods:

• Conducted large-scale experiments with crowd-sourced acceptability judgments.
• Developed an acceptability measure to normalize probability values, controlling for sentence length and lexical frequency.
• Utilized state-of-the-art unsupervised language models from natural language processing for prediction tasks.

Main Results:

• Demonstrated that gradience is a pervasive feature of acceptability judgments.
• Achieved high accuracy in predicting acceptability judgments using augmented probabilistic models.
• Showed strong correlations between model-derived acceptability scores and human judgments.

Conclusions:

• Linguistic knowledge appears to be intrinsically probabilistic.
• Augmented probabilistic models offer a viable approach to predicting sentence acceptability.
• Findings support a probabilistic view of grammatical competence over a strictly binary one.